Π-NAS

This repository provides the evaluation code of our submitted paper: Pi-NAS: Improving Neural Architecture Search by Reducing Supernet Training Consistency Shift.

Our Trained Models

• Here is a summary of our searched models:

  ImageNet	FLOPs	Params	Acc@1	Acc@5
  Π-NAS-cls	5.38G	27.1M	81.6%	95.7%

  Mask-RCNN on COCO 2017	APbb	APmk
  Π-NAS-trans	44.07	39.50

  DeeplabV3 on ADE20K	pixAcc	mIoU
  Π-NAS-trans	81.27	45.47

  DeeplabV3 on Cityscapes	mIoU
  Π-NAS-trans	80.70

Usage

1. Requirements

• Install third-party requirements with command pip install -e .
  • We adapt the code from PyTorch-Encoding and detectron2 to validate our models.
• Prepare ImageNet, COCO 2017, ADE20K and Cityscapes datasets
  • Our data paths are at /data/ImageNet, /data/coco, /data/ADEChallengeData2016 and /data/citys, respectively.
  • You can specify COCO's data path through environment variable DETECTRON2_DATASETS and others in experiments/recognition/verify.py, encoding/datasets/ade20k.py and encoding/datasets/cityscapes.py.
• Download our checkpoint files

2. Evaluate our models

• You can evaluate our models with the following command:

  ImageNet	FLOPs	Params	Acc@1	Acc@5
  Π-NAS-cls	5.38G	27.1M	81.6%	95.7%

  python experiments/recognition/verify.py --dataset imagenet --model alone_resnest50 --choice-indices 3 0 1 3 2 3 1 2 0 3 2 1 3 0 3 2 --resume /path/to/PiNAS_cls.pth.tar

  Mask-RCNN on COCO 2017	APbb	APmk
  Π-NAS-trans	44.07	39.50

  DETECTRON2_DATASETS=/data python experiments/detection/plain_train_net.py --config-file experiments/detection/configs/mask_rcnn_ResNeSt_50_FPN_syncBN_1x.yaml --num-gpus 8 --eval-only MODEL.WEIGHTS /path/to/PiNAS_trans_COCO.pth MODEL.RESNETS.CHOICE_INDICES [3,3,3,3,1,1,3,3,3,0,0,1,1,0,2,1]

  DeeplabV3 on ADE20K	pixAcc	mIoU
  Π-NAS-trans	81.27	45.47

  python experiments/segmentation/test.py --dataset ADE20K --model deeplab --backbone alone_resnest50 --choice-indices 3 3 3 3 1 1 3 3 3 0 0 1 1 0 2 1 --aux --se-loss --resume /path/to/PiNAS_trans_ade.pth.tar --eval

  DeeplabV3 on Cityscapes	mIoU
  Π-NAS-trans	80.70

  python experiments/segmentation/test.py --dataset citys --base-size 2048 --crop-size 768 --model deeplab --backbone alone_resnest50 --choice-indices 3 3 3 3 1 1 3 3 3 0 0 1 1 0 2 1 --aux --se-loss --resume /path/to/PiNAS_trans_citys.pth.tar --eval

Training and Searching

This reimplementation is based on OpenSelfSup and MoCo. Please acknowledge their contribution.

cd OpenSelfSup && pip install -v -e .

1. Π-NAS Learning

bash tools/dist_train.sh configs/pinas_learning.py 8 --work_dir /path/to/save/logs/and/models

2. Extract supernet backbone weights

python tools/extract_backbone_weights.py /checkpoint/of/1. /extracted/weight/of/1.

3. Linear Training

bash tools/dist_train.sh configs/pinas_linear_training.py 8 --pretrained /extracted/weight/of/1. --work_dir /path/to/save/logs/and/models

4. Linear Evaluation

bash tools/dist_train.sh configs/pinas_linear_evaluation.py 8 --resume_from /checkpoint/of/3. --work_dir /path/to/save/logs/and/models